Abstract
This paper presents a robust, language-independent approach to story segmentation in multilingual broadcast news. By utilizing phoneme lattices and expected n-gram counts within a TextTiling framework, the method avoids the limitations of traditional word-based recognition, such as language dependency and out-of-vocabulary (OOV) issues. Experiments on Mandarin and English corpora demonstrate that lattice-based features significantly outperform 1-best phoneme sequences and even exceed word-based baselines in cross-language scenarios, proving the efficacy of probabilistic phonetic information for topic boundary detection.
Introduction
Automatic story segmentation—the process of dividing a continuous stream of audio or text into topically consistent segments—is a fundamental prerequisite for information retrieval and summarization. Traditionally, this task relies on Large Vocabulary Continuous Speech Recognition (LVCSR) to generate word transcripts. However, LVCSR systems are computationally intensive and heavily dependent on language-specific linguistic resources like dictionaries and grammars. Furthermore, they struggle with OOV terms, such as new names or technical jargon common in news broadcasts.
To address these challenges, this research explores a "language-independent" solution using phonemes. Phonemes are the smallest units of sound and are finite in number, often shared across languages, and do not require a massive vocabulary. The authors propose shifting the focus from words to phoneme lattices to preserve the statistical signatures of topics even in noisy environments.
Methodology: Beyond Words to Phoneme Lattices
The core of the proposed system is an adaptation of the classic TextTiling algorithm. TextTiling typically detects topic shifts by measuring the "lexical cohesion" (vocabulary overlap) between adjacent blocks of text. When the overlap drops significantly, a boundary is identified.
The innovation here lies in representing lexical units as phoneme lattices rather than simple word or phoneme strings. A lattice is a directed graph representing multiple possible hypotheses and their acoustic probabilities.
Instead of relying on the recognizer's single best guess (the "1-best" sequence), which is prone to error, the system calculates the expected frequency of phoneme n-grams (sequences of 1 to 5 phonemes) across all possible paths in the lattice. These expected counts provide a richer, more stable statistical signature. These signatures are then compared using a cosine similarity formula, and a "valley-finding" procedure on the resulting scores identifies the story boundaries.
Experimental Evaluation
The system was evaluated using the TDT2 (Topic Detection and Tracking 2) corpora, which includes 53 hours of VOA Mandarin and 111 hours of VOA English broadcast news. The researchers compared their lattice-based approach against manual annotations, word-based LVCSR baselines, and 1-best phoneme sequences.
A key aspect of the study was testing cross-language robustness: using a phoneme recognizer trained on one language (e.g., Mandarin) to segment news in another (e.g., English).
Key Results and Insights
The experimental results highlighted several critical advantages of the lattice-based approach:
- Lattices vs. 1-Best: On the Mandarin corpus, the 3-gram phoneme lattice approach yielded a 3.13% absolute gain in F1-measure over the 1-best phoneme sequence. By accounting for uncertainty, the lattice preserves the statistical "scent" of a topic even when the recognizer is unsure of specific sounds.
- Outperforming Word Baselines: Remarkably, in the English corpus, the Mandarin-trained phoneme lattice system achieved an F1-measure of 0.4095, outperforming the English word-based baseline (0.3669) derived from LVCSR transcripts.
- The N-gram "Sweet Spot": While unigrams lack discriminative power, 3-grams and 4-grams act as effective "pseudo-words," providing enough detail to distinguish topics while remaining robust to recognition noise.
Conclusion and Future Work
The study demonstrates that phoneme lattices are a powerful tool for multilingual story segmentation, offering a path toward language-agnostic systems that are robust to the errors of speech recognition. While performance drops as n-grams become too long (due to data sparsity), the lattice approach consistently outperforms 1-best methods.
Future research will focus on moving toward the International Phonetic Alphabet (IPA) to create a truly universal phoneme recognizer and exploring more advanced segmentation algorithms, such as Normalized Cuts, to further leverage the rich information contained within phoneme lattices.
Summary updated